LawrenceC comments on Sparse autoencoders find composed features in small toy models